Available online at www.sciencedirect.com
ScienceDirect Procedia CIRP 33 (2015) 327 – 332
9th CIRP Conference on Intelligent Computation in Manufacturing Engineering - CIRP ICME '14
Sensor Monitoring during Tack Welding of Aerospace Components Piera Centobellia,b*, Roberto Tetia,b, Leif A. Andersenc a
Dept. of Chemical, Materials and Industrial Production Engineering, University of Naples Federico II, P.le Tecchio 80, Naples, Italy b Fraunhofer Joint Laboratory of Excellence on Advanced Production Technology (Fh-J_LEAPT), P.le Tecchio 80, Naples, Italy c GKN Aerospace Norway AS, Kirkegårdsveien 45, 3616 Kongsberg, Norway
* Corresponding author. Tel.: +390817682371; fax: +390817682362. E-mail address:
[email protected]
Abstract In the 21st century, a growing attention is being paid to the implementation of sensor monitoring systems for innovative production environments based on the conviction that monitoring and control of manufacturing processes is an essential issue to fabricate products with very high quality. In an effort towards zero defect manufacturing concept implementation, an on-line sensor system was developed by incorporating modern technologies of 2D laser distance measurements for process monitoring and control during tack welding of nickel base alloy aerospace components. The system extracts relevant sensorial features from laser line position and intensity signals and integrates them to provide one sensor fusion pattern feature vector to be fed to cognitive paradigms for decision making on desired tack welding process outcome.
© 2014 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license © 2014 The Authors. Published by Elsevier B.V. (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of the International Scientific Committee of “9th CIRP ICME Conference". Selection and peer-review under responsibility of the International Scientific Committee of “9th CIRP ICME Conference” Keywords: Sensor monitoring; Feature extraction; Compliant assembly; Tack welding; Nickel superalloy.
1. Introduction Numerous studies have shown that the success of modern flexible manufacturing systems is largely based on the availability of real-time data about process operating conditions (materials, tools, processess) [1-6]. Reliable models of manufacturing performance prediction are, in many cases, very difficult to find. Therefore, to avoid manufacturing errors, defects and malfunctions, methodologies for process monitoring and control are highly desirable. The use of advanced sensing devices for process monitoring can decidedly foster production systems to achieve optimal performance, high availability of production resources, and excellent final product quality [6]. It can speed up automation extension, including automation of complex and costly manual operations, and contribute to deeper process understanding by process data capture and storage for later analysis [7]. To increase the effectiveness of monitoring systems, improved sensor technology and innovative signal processing techniques need to be applied. The design of intelligent sensor monitoring systems with advanced paradigms of signal processing for cognitive decision making allows to obtain
more comprehensive information about processing conditions, with the aim of increasing economic efficiency through fully optimized process control. For a given process monitoring application, one sensor signal type may not be able to meet all the required performance specifications [6]. The use of multiple sensor units of different nature permits the combination of sensory data from disparate sources so that the resulting information are more accurate and complete than would be feasible when the sources are used separately [8]. In such a case, sensor fusion technology can be used to achieve the solution of the monitoring problem [9]. Monitoring of welding processes saw the first applications in the ’90s. In [10] a device was developed for real-time monitoring of infrared emissions during laser welding. Many other sensor monitoring techniques have been implemented since in order to provide real-time welding process control and assure high quality weldments [11]. These techniques include acoustic emission methods [12-14], optical techniques [15,16], and thermal sensing methods [17]. Spot welding is a discontinuous welding process that can be critical in the case of difficult to weld materials with properties negatively influencing the weld quality [18]. This stimulates the use of sensor monitoring methodologies, such
2212-8271 © 2014 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Selection and peer-review under responsibility of the International Scientific Committee of “9th CIRP ICME Conference” doi:10.1016/j.procir.2015.06.072
328
Piera Centobelli et al. / Procedia CIRP 33 (2015) 327 – 332
as in [19,20] where a magnetic sensor array [19] and a multiple-sensor system [20] and were used for real-time monitoring applied to spot welding of aluminium alloy components. Spot welding is particularly widespread in the automotive industry where most of the existing sensor monitoring systems are based on current, voltage, and welding force signal detection and analysis [21]. Furthermore, the use of high-performance monitoring systems based on digital CMOS sensors and image processing technologies is rapidly expanding [22]. Tack welding is a spot welding process that generates a temporary kind of weld utilised during assembly to fix the position of the two metal parts to be later fully welded [18]. The temporary nature of tack welding does not make it less important as it represents the starting point of the complete welding process and a mistake at this stage determines an irreversible error with critical impact on the final welded product quality. In this paper, a 2D laser sensor system has been employed for real-time monitoring of tack welding carried out on nickel base alloy aerospace components. 2. Tack welding process This research has been carried out in collaboration with GKN Aerospace Engine Systems Norway AS, an aerospace company producing high quality jet engine components for, amongst other programs, GEnx 1B and GEnx 2B engines for the B 787 Dreamliner and the new version of the B747. The work had the main scope to realise an on-line sensor monitoring system for the self-adjustment of the welding process parameters in full seam welding of turbine rear frame (TRF) segments made of nickel superalloy (Fig. 1). Welding takes place between a panel, which is the moving part, and a strut, which is the fixed part. Before starting the welding operation, the struts are initially mounted on a fixture and then an assembly robot takes a panel and positions it between two struts to allow for tack welding (Fig. 2).
Fig. 2. Assembly and welding robot.
During the alignment between panel and strut, the detection of the surface offset between the two parts is a critical issue. The panel is moved by the robot into place within the goal tolerance based on the measurement of this surface offset. The tolerances for the panel and strut are more than 3 times larger than that of the allowed final offset on the TRF. Once the panel position is correctly achieved, the first tack weld is performed on the centre of the panel and strut assembly. The panel is thus locked in position relative to the strut with the tack weld and the attention is moved to a different position of the strut/panel interface. The measurement is repeated but, after the initial placement, the panel is bent within the surface offset tolerance. The measuring, bending, and tack welding operations are repeated above and below the central tack weld (~ 20 times over a seam length of ~ 250 mm) until the seam line is fully tacked using a rotation axis to move the next strut/panel interface towards the robots. The tack welding procedure is reiterated for all strut/weld interfaces of the TRF which, at completion, is ready for final/finishing tungsten inert gas (TIG) arc welding process. A vital quality criterion for tack welding is that the mating surfaces must not be separated by more than the allowable tolerance. The assembly process is performed by skilled human operators to ensure an allowable surface offset between panel and strut. In this manual process, tack welding is carried out by TIG welding with quite large tacks (4 to 5 mm diameter) subjectively distributed by the human operator along the weld seam line. By automating the surface offset measurement, the robot, provided with feedback from an offset sensor, performs the manipulation of the plate and strut to be tack welded so that optimal weldments are realised. 3. Sensor monitoring system
Fig. 1. Exploded view of the 1.5 m diameter TRF. Panels (dark red) and struts (brown).
The final aim of this research work is to realize a monitoring system for self-adjustment of welding process parameters during full seam welding of aerospace components consisting of TRF segments made of nickel superalloy. As previously mentioned, the parts to be welded are the TRF panels and struts. The relative position of the each panel and strut couple, prior to tack welding, is extremely important to satisfy the final welded product quality requirements. In this paper, a sensor fusion technology approach is implemented to monitor the compliant assembly and tack welding processes and support decision makers in order to
329
Piera Centobelli et al. / Procedia CIRP 33 (2015) 327 – 332
evaluate whether the relative position of the parts is within tolerance before full seam welding is performed. The position measurements refer to the two mating parts, plate and strut, and must be carried out on an on-line basis. Because of the high number of measurements needed for just one seam weld, measuring must be comparatively fast which means that no contact metrology method is acceptable. Accordingly, a 2D laser sensor system was selected to carry out the surface offset measurements for tack welding process monitoring. Fig. 2 illustrates the welding setup and the optical instrumentation setup for 2D laser line measurements during the assembly and tack welding processes. This procedure, though currently carried out manually, can be easily automated for industrial implementation. During the alignment of the panel with the strut, a laser line is projected by the laser sensor onto the centre of both panel and strut to detect the surface offset between them, as shown in Fig. 3. The sensor system acquires height images and profiles by means of a laser sheet-of-light triangulation technique: a laser line is projected onto the part, the resulting height image is detected by the camera and converted into a height profile. The laser line signal acquisition is at present carried out manually and, accordingly, the scanning frequency is subjectively decided. For each scan, the sensor outputs two *.txt files representing the laser line position signal and the laser line intensity signal. These output signals constitute the input to the MatLab® features extraction procedure developed to monitor the relative parts position before tack welding.
(a)
4. Signals processing for features extraction The positioning of the TRF segments for the welding operation is controlled by monitoring the surface offset between them. Thus, the purpose of signal processing is to evaluate the surface offset at the weld seam in order to optimally adjust the relative parts position.
(b) Fig. 4. (a) Detection of the laser line intensity step drop position; (b) identification of minimum intensity value within the step drop.
The experimental sensorial data set was made available by NTNU [23] comprising a total of 4050 signals, i.e. 2025 laser line position signals plus the corresponding 2025 laser line intensity signals, detected during the performance of one tack weldment. From both these signal types, relevant signal features need to be selected and extracted. The processing of the laser line signals was realised through MatLab® by zooming in on the plotted signals to identify the exact position of the step drop (Fig. 4a) and the minimum intensity value within the step drop (Fig. 4b). 4.1. Laser line position features
Fig. 3. Laser line projected on the centre of panel and strut assembly to detect the surface offset between them. Laser distance sensor model C4-2040-GigE by Automation Technology.
The laser line position features were selected on the basis of the real instances of laser line signals experimentally provided in the form of 2025 subsequent laser line position signals detected during one full panel displacement towards the fixed strut to reach the final position for one tack welding. An example of laser line position signal is shown in Fig. 5. A total of 10 features (F1 - F10) to be extracted from each laser line position signal were identified based on their usefulness for assembly and tack welding process monitoring.
330
Piera Centobelli et al. / Procedia CIRP 33 (2015) 327 – 332 4
4.5
x 10
4
height [a.u.]
3.5
height [a.u.]
3 2.5
2
1.5 1
position [px]
0.5
0
0
500
1000 1500 Position [px]
2000
2500
position [px]
Fig. 6. Representation of right-most point of panel (red), left-most point of the strut (green), middle line position, segment gap, and surface offset width ΔY between panel and strut.
The first feature (F1) is the sequential number of the analysed laser line position signal. The further 9 features (F2 F10), described below, were extracted from each laser line position signal through MatLab® scripts purposely created: x Abscissa and ordinate of the right-most point of the panel, F2 and F3; x Abscissa and ordinate of the left-most point of the strut, F4 and F5; x Middle line position, F6; x Segment gap, F7; x Surface offset ΔY between panel and strut, F8; x Left angle (degrees), F9; x Right angle (degrees), F10. The panel is the moving part, so the abscissa and the ordinate of the right-most point of the panel will change for each subsequent laser line position signal. For each laser line position signal, the right-most point of the panel (with coordinates F2, F3) and the left-most point of the strut (with coordinates F4, F5) are identified as shown in Fig. 6 where they are marked in red and green, respectively. The middle line position (F6) is given by the midpoint of the segment gap (F7), the latter being the distance (measured parallel to the abscissa axis) between the right-most point of the panel and the left-most point of the strut for each laser line position signal. If the panel is very close to the strut, the offset tends to zero or if the panel has been bent towards the strut, the right-most point of the panel and the left-most point of the strut tend to superimpose and coincide also with the middle line position. Fig. 6 shows an example of middle line position and segment gap. The surface offset (F8) between panel and strut is evaluated as the difference between the left-most point ordinate of the panel and the right-most point ordinate of the panel, as shown in Fig. 6. For all laser line position signals, the inclination angles of both panel and strut (F9, F10), respectively indicated with ThetaLL and ThetaLR, were evaluated as shown in Fig. 7. To evaluate the ThetaLL inclination angle, the exact location of the left and right endpoints of segment LL must be identified (Fig. 7).
height [a.u.]
Fig. 5. Laser line position signal vs. position in pixel.
position [px]
Fig. 7. Inclination angles representation.
To determine the LL left endpoint, the values of the Edge Offset (30 px) and the LL dimension (64 px), that are predefined data, were subtracted from the panel right-most point abscissa in px. As regards the identification of the LL right endpoint, only the value of the Edge Offset was subtracted from the panel right-most point abscissa in px. Thus, a vector of length equal to LL, represented in green in Fig. 8, was obtained by segmenting the original laser line position signal and considering the abscissa values of its extremities. Then, the MatLab® function ‘polyfit’ was used to find the coefficients of a 1st degree polynomial function that best fits the data points by a least-squares regression. The first coefficient is the slope of the regression straight line, represented in red in Fig. 8, which is equal to the trigonometric tangent of angle ThetaLL. To obtain the positive ThetaLL angle value in degrees, the arctangent of the slope was calculated and multiplied by -(180/π). Similarly, to evaluate the ThetaLR inclination angle, the exact location of the left and right endpoints of segment LR must be identified as seen before. A vector of length equal to LR was obtained by segmentation of the original laser line and the MatLab® function ‘polyfit’ was used to find the coefficients of a 1st degree polynomial function that best fits the data point by a least-squares regression (Fig. 9).
331
Piera Centobelli et al. / Procedia CIRP 33 (2015) 327 – 332 10000 9000
intensity Intensity [a.u.]
height [a.u.]
8000
position [px]
7000 6000 5000 4000 3000 2000
Fig. 8. Representation of segment LL (green) and best fit regression line (red).
1000 0
0
500
1000
1500
2000
2500
Position [px]
position [px]
height [a.u.]
Fig. 10. Laser line intensity signal plotted vs. position through MatLab®.
position [px]
Fig. 9. Representation of segment LR (green) and best fit regression line (red).
The first coefficient is the slope of the regression straight line represented in red in Figure 9, that is equal to the trigonometric tangent of the angle ThetaLR. To obtain the positive ThetaLR angle value in degrees, the arctangent of the slope was calculated and multiplied by (180⁄π). 4.2. Laser line intensity features The laser line intensity features were selected on the basis of the real instances of laser line signals experimentally provided in the form of 2025 subsequent laser line intensity signals detected during one full panel displacement towards the fixed strut to reach the final position for one tack welding. Fig. 10 shows an example of laser line intensity signal. A total of 4 features (F11 - F14) to be extracted from the laser line intensity signals were identified based on their usefulness for the monitoring of assembly and tack welding processes. The four features, described below, were extracted from each laser line intensity signal through MatLab® scripts purposely created: x Position (abscissa) and value (ordinate) of the signal global minimum, F11 and F12; x Position (abscissa) and value (ordinate) of the minimum in the signal drop window, F13 and F14: 5. Sensor fusion technology approach To improve the knowledge extraction from the laser line position and intensity signals for enhanced monitoring of TRF segments tack welding, a sensor fusion technology approach was implemented to synergically combine the diverse features extracted from the two signal types.
Fig. 11. Sensor fusion concept for the test case and sensor fusion pattern vector construction procedure.
Based on the two procedures previously developed for feature extraction from each of the two laser line signals (position and intensity), a MatLab® sensor fusion code was implemented that is capable to receive in input the sensor signal couples and generate in output one sensor fusion pattern feature vector comprising the combination of the features extracted from both signal types. Fig. 11 shows that the pattern feature vector output of the sensor fusion data processing is a matrix with 14 columns, one for each feature, and 2025 rows, one for each couple of laser line position signal and corresponding laser line intensity signal. In this way, the sensor fusion signal processing methodology provides sensor fusion pattern feature vectors ready to be fed to cognitive paradigms for real-time robust decision making on tack welding process quality. 6. Conclusions and future developments The main purpose of this paper is the implementation of a sensor fusion technology approach to generate sensor fusion pattern feature vectors synergically combining the features extracted from two laser line signal types, i.e. position and intensity signals, for on-line sensor monitoring during tack welding of nickel base alloy aerospace components.
332
Piera Centobelli et al. / Procedia CIRP 33 (2015) 327 – 332
The knowledge content of the sensor fusion pattern feature vectors can be assessed in terms of success rate in the identification of correct parts positioning for the assembly and tack welding processes. This can be realised by developing cognitive decision making paradigms for pattern recognition (e.g. neural network, fuzzy logic, neuro-fuzzy systems) to find correlations between the input sensor fusion pattern feature vectors and the output quality parameters, represented by the correct panel and strut relative position before tack welding, indispensable to determine the acceptability of the process and the quality of the welded product. Acknowledgments This research work has received funding support from the EC FP7 Integrated Project (IP) on Intelligent Fault Correction and self Optimizing Manufacturing systems (IFaCOM) under grant agreement n. 285489. Moreover, Lars Tingelstad of NTNU, partner in the IFaCOM project, is thankfully accredited for supplying the sensorial data set. The Fraunhofer Joint Laboratory of Excellence for Advanced Production Technology (Fh-J_LEAPT Naples) at the Department of Chemical, Materials and Industrial Production Engineering, University of Naples Federico II, is gratefully acknowledged for its support to this work. References [1] Wen X-Q, Zhao Z-M, Xie Z-Y. Integrated study of condition monitoring and fault diagnosis system for flexible manufacturing system. Advanced Materials Research; 2011; 179:180, p. 678-684. [2] Zhao YF, Xu X. Enabling cognitive manufacturing through automated on-machine measurement planning and feedback. Advanced Engineering Informatics; 2010; 24:3, p. 269-284. [3] Reilly T. A review of signal processing and analysis tools for comprehensive rotating machinery diagnostics, Conference Proceedings of the Society for Experimental Mechanics Series; 2011; 5, p. 463-479. [4] Rubio EM, Teti R. Cutting parameters analysis for the development of a milling process monitoring system based on audible energy sound. J. of Intelligent Manufacturing; 2009; 20:1, p. 43-54. [5] Mhalla A, Jenhani O, Dutilleul SC, Benrejeb M. Contribution to the monitoring of manufacturing systems with time constraints: Application to a surface treatment line. 14 th Int. Conf. on Sciences and Techniques of Automatic Control and Computer Engineering (STA); 2013; p. 243-250.
[6] Teti R, Jemielniak K, O’Donnell G, Dornfeld D. Advanced Monitoring of Machining Operations. CIRP Annals; 2010; 59:2, p. 717–739. [7] EC FP7 FoF-NMP Integrated Project (IP) on Intelligent Fault Correction and self Optimizing Manufacturing systems – IFaCOM, www.ifacom.it. [8] Hall DL, Llinas J. Handbook of multisensor data fusion. CRC Press LLC, New York, 2001. [9] Simeone A, Segreto T, Teti R. Sensor Fusion for Tool State Classification in Nickel Superalloy High Performance Cutting. Procedia CIRP; 2012; 1, p. 593-598. [10] Sanders PG, Keske JS, Kornecki G, Leong KH. Real-time monitoring of laser beam welding using infrared weld emissions. J. of Laser Appl.; 1998; 10, p. 205. [11] Kah P, Layus P, Hiltunen E, Martikainen J. Real-time weld process monitoring. 3rd Int. Conf. on Manufacture Engineering, Quality and Production System (ICMEQP); 2014; Hong Kong; China; 27-28/2/ 2014. [12] Hamann C, Rosen HG, Labiger B. Acoustic emission and its application to laser spot welding. Conf. on High Power Lasers and Laser Machining Technology (SPIE); 1989. [13] Weeter L, Albright C. The effect of full penetration on laser induced stress wave emissions during laser spot welding. Materials Evaluation; 1987; 45, p. 353–357. [14] Jon MC. Non-contact acoustic emission monitoring of laser beam welding. Welding Journal; 1985; 64, p. 43–48. [15] Chen HB, Li L, Steen WM, Brookfield DJ. Multi-frequency fibre optic sensors for in-process laser welding quality monitoring. J. of NonDestructive Testing and Evaluation; 1993; 26:2, p. 67–73. [16] Juptner W, Rothe R, Sepold G. Mirror optics for high-power laserbeams, 6th Int. Symp. on Gas Flow and Chemical Lasers; Jerusalem; September 1986; Berlin, Springer. [17] Webber T. Real time laser welding quality control system. 6th Int. Congr. on Applications of Lasers and Electro-Optics (ICALEO’87); San Diego; 1987; Berlin, Springer. [18] Cary HB, Helzer SC. Modern Welding Technology. 6 th ed. Pearson/Prentice-Hall, Upper Saddle River, N. J., 2005. [19] Wang R, Luo Z, Li Y, Xuan W, Fan N. Modified signal processing method for spot welding monitor system. Proc. SPIE 7997, 4th Int. Seminar on Modern Cutting and Measurement Engineering; 2011. [20] Pan C, Qin P. Multiple-sensor monitor resistance spot welding process. Int. Conf. on E-Product, E-Service and E-Entertainment (ICEEE); Henan; 2010, p. 1-4. [21] Cullen JD, Athi N, Al-Jader M, Johnson P, Al-Shamma’a AI, Shaw A, El-Rasheed AMA. Multisensor fusion for on line monitoring of the quality of spot welding in automotive industry. Measurement; 2008; 41:4, p. 412–423. [22] Fang J, Chen Y, Li L, Wu L. Coaxial monitoring with a CMOS camera for CO2 laser welding. Proceedings of SPIE - The International Society for Optical Engineering; 2005; 13, p. 101-109. [23] Tingelstad L, Egeland O. Robotic assembly of aircraft engine components using a closed-loop alignment process. Procedia CIRP; 2014; 23, p. 110-115.
